app.py

from typing import Tuple, Dict
import requests
import random
import numpy as np
import gradio as gr
import torch
from PIL import Image
from diffusers import StableDiffusionInpaintPipeline

INFO = """
# FLUX-Based Inpainting 🎨

This interface utilizes a FLUX model variant for precise inpainting. Special thanks to the [Black Forest Labs](https://huggingface.co/black-forest-labs) team 
and [Gothos](https://github.com/Gothos) for contributing to this advanced solution.
"""

# Constants
MAX_SEED_VALUE = np.iinfo(np.int32).max
TARGET_DIM = 1024
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# Function to clear background
def clear_background(image: Image.Image, threshold: int = 50) -> Image.Image:
    image = image.convert("RGBA")
    pixels = image.getdata()
    processed_data = [
        (0, 0, 0, 0) if sum(pixel[:3]) / 3 < threshold else pixel for pixel in pixels
    ]
    image.putdata(processed_data)
    return image

# Sample data examples
EXAMPLES = [
    [
        {
            "background": Image.open(requests.get("https://example.com/doge-1.png", stream=True).raw),
            "layers": [clear_background(Image.open(requests.get("https://example.com/mask-1.png", stream=True).raw))],
            "composite": Image.open(requests.get("https://example.com/composite-1.png", stream=True).raw),
        },
        "desert mirage",
        42,
        False,
        0.75,
        25
    ],
    [
        {
            "background": Image.open(requests.get("https://example.com/doge-2.png", stream=True).raw),
            "layers": [clear_background(Image.open(requests.get("https://example.com/mask-2.png", stream=True).raw))],
            "composite": Image.open(requests.get("https://example.com/composite-2.png", stream=True).raw),
        },
        "neon city",
        100,
        True,
        0.9,
        35
    ]
]

# Load model
inpainting_pipeline = StableDiffusionInpaintPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16).to(DEVICE)

# Utility to adjust image size
def get_scaled_dimensions(
    original_size: Tuple[int, int], max_dim: int = TARGET_DIM
) -> Tuple[int, int]:
    width, height = original_size
    scaling_factor = max_dim / max(width, height)
    return (int(width * scaling_factor) // 32 * 32, int(height * scaling_factor) // 32 * 32)

@spaces.GPU(duration=100)
def generate_inpainting(
    input_data: Dict,
    prompt_text: str,
    chosen_seed: int,
    use_random_seed: bool,
    inpainting_strength: float,
    steps: int,
    progress=gr.Progress(track_tqdm=True)
):
    if not prompt_text:
        return gr.Info("Provide a prompt to proceed."), None

    background = input_data.get("background")
    mask_layer = input_data.get("layers")[0]

    if not background:
        return gr.Info("Background image is missing."), None

    if not mask_layer:
        return gr.Info("Mask layer is missing."), None

    new_width, new_height = get_scaled_dimensions(background.size)
    resized_background = background.resize((new_width, new_height), Image.LANCZOS)
    resized_mask = mask_layer.resize((new_width, new_height), Image.LANCZOS)

    if use_random_seed:
        chosen_seed = random.randint(0, MAX_SEED_VALUE)

    torch.manual_seed(chosen_seed)
    generated_image = inpainting_pipeline(
        prompt=prompt_text,
        image=resized_background,
        mask_image=resized_mask,
        strength=inpainting_strength,
        num_inference_steps=steps,
    ).images[0]

    return generated_image, resized_mask

# Build the Gradio interface
with gr.Blocks() as flux_app:
    gr.Markdown(INFO)

    with gr.Row():
        with gr.Column():
            image_editor = gr.ImageEditor(
                label="Edit Image",
                type="pil",
                sources=["upload", "webcam"],
                brush=gr.Brush(colors=["#FFF"], color_mode="fixed")
            )

            prompt_box = gr.Text(
                label="Inpainting Prompt", placeholder="Describe the change you'd like."
            )
            run_button = gr.Button(value="Run Inpainting")

            with gr.Accordion("Settings"):
                seed_slider = gr.Slider(0, MAX_SEED_VALUE, step=1, value=42, label="Seed")
                random_seed_toggle = gr.Checkbox(label="Randomize Seed", value=True)
                inpainting_strength_slider = gr.Slider(0.0, 1.0, step=0.01, value=0.85, label="Inpainting Strength")
                steps_slider = gr.Slider(1, 50, step=1, value=25, label="Inference Steps")

        with gr.Column():
            output_image = gr.Image(label="Output Image")
            output_mask = gr.Image(label="Processed Mask")
            
    run_button.click(
        generate_inpainting,
        inputs=[image_editor, prompt_box, seed_slider, random_seed_toggle, inpainting_strength_slider, steps_slider],
        outputs=[output_image, output_mask]
    )

    gr.Examples(
        examples=EXAMPLES,
        fn=generate_inpainting,
        inputs=[image_editor, prompt_box, seed_slider, random_seed_toggle, inpainting_strength_slider, steps_slider],
        outputs=[output_image, output_mask],
        run_on_click=True,
    )

flux_app.launch(debug=False, show_error=True)